Guess your size: A hybrid model for footwear size recommendation

Abstract

In recent years, online shopping for footwear has rapidly increased. However, the user experience has not been satisfactory because of the size mismatch problem, i.e., customers often fail to choose the right size online. Traditional size selection schemes, including those suggesting that users select footwear sizes according to their past experiences or those based on simple measurements, usually result in a high return rate of up to $35\%$. The limitation of the traditional size selection schemes is that they fail to consider (1) the characteristics of foot shapes and (2) the preferences of individual customers. In this paper, we propose a size recommendation framework that is jointly based on 3D (foot and last) features and user preference. First, we report measurement studies of foot shape characteristics based on foot data for 10 K individuals. Our findings reveal that users have diverse foot shapes and different personal preferences regarding size matching. Second, based on our measurement insights, we design a size recommendation model that jointly considers 3D foot models, shoe characteristics and user preferences. We also provide a predictive model that predicts comfort levels for particular parts of the foot based on the given size recommendation. Finally, our data-driven experiments show that the proposed size recommendation improves the size selection accuracy to $92\%$, which is a $22\%$ improvement compared to conventional solutions.

Introduction

Recent years have witnessed a rapid growth in online shopping for footwear. Smart Company Australia’s website reports that online shoe sales revenue grew by $17.4\%$ from 2009 to 2014.¹ However, due to mismatches in shoe size, the return rate can be up to $35\%$, much higher than that for traditional retail stores.² The main reasons underlying this phenomenon are as follows. First, customers often choose incorrect sizes. According to our measurement studies on over 10 K people (involving a questionnaire-based survey), the traditional size selection process used by online consumers has an accuracy of less than $70\%$ (the accuracy of foot-length-based and regular-size-based recommendations is $33.7\%$ and $69.7\%$, respectively). Second, the sizes offered fail to cover a significant fraction of customers. The measurement results show that as many as $26\%$ of customers may be unable to wear any of the sizes available for any given shoe model. Consequently, size recommendation has become an important target of investigation.

Sizing systems also differ in the units of measure that they use. This difference results in different increments between shoe sizes for shoes with different shoe items because different shoe lasts are often used. Therefore, people seldom wear a fixed shoe size across all shoe styles or even across all brands of the same company, and customers often switch among two or three sizes when choosing pairs of shoes, which makes it difficult to recommend suitable sizes to users.

Data-driven approaches for size recommendation based on large-scale sale/return records, in which size selection is formulated as a user–item recommendation problem that can be solved using well-developed collaborative filtering[1] and matrix factorization solutions[2], are beginning to show promise. However, such methods tend to consider only superficial information concerning user foot and shoe characteristics. Therefore, they fail to provide size recommendations for users with unique foot shapes or personal preferences.

To address these challenges, in this paper, we propose a size recommendation framework that is jointly based on 3D (foot and last³) features and user preferences. We establish a random forest model to obtain the matching structure relationship between feet and lasts. To achieve a better recommendation performance, the model is modified by taking the effect of user preferences into account. Because the selection of suitable sizes alone cannot meet user needs in terms of quality requirements, an investigation of wearing experiences related to different parts of users’ feet is also conducted to further improve size recommendation.

In particular, our contributions are as follows.

• $▷$ First, we report measurement studies on foot shape characteristics based on 3D foot data for 10 K individuals and shoe last information for 24 types of women’s shoes and 8 types of men’s shoes as well as subjective fitting records.

• $▷$ Second, based on our measurement insights, we design a feature selection strategy to extract features from both 3D foot models and shoe last data based on expert experience. We capture users’ size preferences by analyzing the user differences in shoe sizes (e.g., some users prefer a larger-than-typical size). Based on these features, we design a recommendation model that jointly considers 3D features and user preferences.

• $▷$ Third, we propose a deep learning model to predict comfort levels for different parts of the foot based on a given size recommendation. With these results, users can make more informed decisions.

The remainder of the paper is organized as follows. We discuss related work in Section 2. We present our data-driven measurement methodology and size recommendation design in Section 3. We evaluate the effectiveness of our design in Section 4. Finally, we conclude the paper in Section 5.